Why does “every page looks fine alone, but the full-site audit scores poorly” happen?
The most common reaction from clients who get a low score on their first full-site audit:
“I manually checked 5 random pages, and every one of them is well written. So why does the whole site only get 62?”
The answer: AI doesn’t evaluate a website on its average score — it does sampling + consistency checking.
In other words:
- It randomly samples 30 pages
- It checks whether those 30 pages are “uniform in quality,” “consistent in structure,” and “convergent in topic”
- It samples 5 great pages but 25 poor ones → the whole site gets judged at the level of those 25
- It samples 30 mediocre pages → the whole site is judged mediocre
- It samples 5 great pages + 25 with structural problems → whole-site downrank
This logic is called the site-level consistency penalty. Below we break down the three most common patterns.
Pattern 1: Template sameness (most common on e-commerce / catalog sites)
The scenario
You have 100 product pages. On every page: - The “How to use” section is nearly identical (only the product name changes) - The “Brand story” section is the same across all 100 pages - The FAQ uses the same template
Each page on its own: has product schema, has images, has price — looks compliant.
What AI sees after sampling the whole site: a 70%+ content duplication rate across those 100 pages.
How AI interprets it
AI doesn’t know whether you “have to keep brand narrative consistent” or whether you’re “padding with templates.” Its heuristic is:
“A normal website should have unique content on every page; 70%+ content overlap = low content productivity, possibly SEO manipulation, possibly AI-generated, possibly a dead site.”
→ Every dimension tied to that score (content citability, E-E-A-T, signs of AI code generation) gets downranked across the entire site.
The fix
Don’t grind through “rewriting all 100 pages from scratch.” Instead:
Split into three layers
- Shared baseline information (product price, spec sheet) → keep it identical, that’s fine
- Unique core content (usage case studies, customer testimonials, real-world scenarios) → must differ on every page
- Brand messaging (brand story, customer service info) → pull it into the footer or a standalone About page, don’t repeat it on every product page
Rewriting strategy
- Reframe each page’s “How to use” by leading with a specific use scenario:
- ❌ All 100 pages write “Steps for using this product: …”
- ✅ Page A writes “Use scenario for office white-collar workers: …”, Page B writes “Use scenario for outdoor workers: …”
- Add a passage of real customer case study to each page (with concrete, quantified results)
The goal: bring the cross-page duplication rate down from 70% to < 30%.
Pattern 2: Inconsistent bylines (most common on media / blogs)
The scenario
Your blog has 300 articles, distributed as: - 30% have a full author page + byline - 40% only say “Editorial team,” no real name - 30% have no byline at all
Each article on its own: has an H1, a publish date, structure.
What AI sees after sampling the whole site: author-signal coverage below 50%.
How AI interprets it
AI’s logic for evaluating E-E-A-T:
“This site’s content is partly under author accountability, but mostly not. That means the editorial standard isn’t uniform across the site, and credibility can only be taken as the site-wide average.”
→ It even drags down the citation rate of that 30% of bylined articles — because AI evaluates at the site level, not just article by article.
The fix
Don’t accept the half-measure compromise of “filling in half of them”:
Pick one of three
- Site-wide named-author system (ideal): every article has a real-named author + author page
- Site-wide organizational-author system: use a single organization name (“ZTPawn Editorial Team”), but the organization page must be as complete as a personal author page (founding year, editorial standards, contact details, team member photos)
- A hybrid, but with ≥ 80% on one side: the remaining 20% of inconsistency is noise AI tolerates
The worst case is the 30/40/30 three-way split — none of the proportions is strong enough, and the site-wide signal gets diluted.
Order of reinforcement
- Day 1: Inventory the distribution (use a Google Sheet to list each article’s byline status)
- Day 7: Pick one primary route (personal or organizational)
- Day 14: Align the inconsistent 50%+ of articles
- Day 21: Run the audit again — the E-E-A-T dimension usually rises noticeably
Pattern 3: Topic drift (most common on B2B SaaS / multi-product lines)
The scenario
You’re a B2B SaaS company, and your site includes: - Homepage: talks about your core product X - Blog: 50 articles about marketing automation - Case study pages: talk about the technical details of customer data migration - White papers: talk about security compliance
Each section on its own: well written, professionally done.
What AI sees after sampling the whole site: topics scattered across 4–5 domains, making it hard to judge “what field you’re actually an expert in.”
How AI interprets it
AI’s logic for building “topical authority”:
“This site touches 5 unrelated topics, each with < 20 pieces of content depth. None of them reaches the ‘authority’ threshold. On an ‘X tool recommendation’ query, the comparison set will be competitors who focus on that topic.”
→ AI’s probability of recommending you is diluted by topic spread.
The fix
Key principle: for B2B sites, “topic density” beats “topic breadth”
- Better 100 in-depth articles around one core topic than 200 scattered across 5 topics
Reclassify
Reclassify your existing content under 3–4 first-level topics:
First-level topic: Customer Data Management (core)
├─ Subtopic: data migration
├─ Subtopic: data governance
└─ Subtopic: data security
First-level topic: B2B Marketing Automation (secondary core)
├─ Subtopic: lead nurturing
└─ Subtopic: marketing funnel
First-level topic: Customer Success (supporting core)
└─ Subtopic: onboarding design
Content that doesn’t fit this structure (e.g. past pieces on “office anecdotes” or “company culture”) → move it to the /about/ section or cut it.
Internal linking strategy
- Articles on the same topic link to each other (each one proactively links back to 3–5 related-topic articles)
- Breadcrumbs show the topic hierarchy:
Home > Customer Data Management > Data Migration > This Article - A topic index page (hub page) serves as the entry point for each first-level topic
Expected effect
After 3–6 months, AI will start considering you on queries like “[your core topic] + tools.” But the precondition is: you genuinely cut the irrelevant topic sprawl.
How do you self-diagnose which pattern your site falls into?
Run a full-site audit (not a single-page one) and look at the following three signals:
| Signal | Corresponding pattern |
|---|---|
| Cross-page content similarity > 50% | Pattern 1: Template sameness |
| Author-signal coverage between 30–70% | Pattern 2: Inconsistent bylines |
| On-site topics scattered across ≥ 4 first-level categories | Pattern 3: Topic drift |
The three patterns may co-exist (e-commerce commonly has Pattern 1 + Pattern 2; B2B SaaS commonly has Pattern 2 + Pattern 3).
The shared principle for fixes: converge first, then expand
Whatever the pattern, the core of the fix is always:
- Converge first (cut / standardize / unify) — align or remove the inconsistent parts
- Expand later (add / write / accumulate) — produce new content under the converged standard
Many clients’ instinctive reaction is “write another 100 pages to dilute the bad parts” — this only makes the problem bigger, because AI will sample even more bad examples.
Deal with cross-page consistency first, then produce new content. Get the order wrong and your ROI is negative.
Engineering detail: how does the site_level analyzer compute these?
GeoWeb’s full-site audit (the M3-X site_level analyzer) runs the following:
- Cross-page content hash similarity: samples 10–30 pages, computes a sliding-window content hash, and flags duplicates above a threshold
- Author-signal coverage: scans each page’s author meta / Article schema author / visible byline, and computes the share of pages with ≥ 1 type of signal
- Topic clustering: uses simple TF-IDF + internal-link network analysis to estimate the number of first-level topics
- Site-wide average score vs. lowest sampled score gap: a gap greater than 25 points flags “uneven quality”
These numbers are presented in the full-site report as site-level health indicators, not just per-page scores.
First step: run a full-site audit (not just a single page)
👉 Free single-page audit — a starting point to see your per-page score
Full-site analysis (including cross-page consistency / author-signal coverage / topic clustering) is currently a custom analysis included with the GEO consulting service; it requires crawling 30–100 pages and producing a site-level report: [email protected]
If you already have the feeling that “every page looks fine, but the whole site just won’t lift” — it’s most likely at least one of the three patterns in this post. Confirm which one first, then apply the matching fix.
GEO in-depth series. Previous post: «Traditional Chinese / Simplified Chinese / English — how big is the citation-rate gap for the same brand across LLMs?»